Computing and Predicting Winning Hands in the Trick-Taking Game of Klaverjas

Computing and Predicting Winning Hands in the Trick-Taking Game of Klaverjas Jan N. van Rijn2;4, Frank W. Takes3;4, and Jonathan K. Vis1;4 1 Leiden University Medical Center 2 Columbia University, New York 3 University of Amsterdam 4 Leiden University Abstract. This paper deals with the trick-taking game of Klaverjas, in which two teams of two players aim to gather as many high valued cards for their team as possible. We propose an efficient encoding to enumerate possible configurations of the game, such that subsequently αβ-search can be employed to effectively determine whether a given hand of cards is winning. To avoid having to apply the exact approach to all possible game configurations, we introduce a partitioning of hands into 981;541 equivalence classes. In addition, we devise a machine learning approach that, based on a combination of simple features is able to predict with high accuracy whether a hand is winning. This approach essentially mimics humans, who typically decide whether or not to play a dealt hand based on various simple counts of high ranking cards in their hand. By comparing the results of the exact algorithm and the machine learning approach we are able to characterize precisely which instances are difficult to solve for an algorithm, but easy to decide for a human. Results on almost one million game instances show that the exact approach typically solves a game within minutes, whereas a relatively small number of instances require up to several days, traversing a space of several billion game states. Interestingly, it is precisely those instances that are always correctly classified by the machine learning approach. This suggests that a hybrid approach combining both machine learning and exact search may be the solution to a perfect real-time artificial Klaverjas agent. Keywords: trick-taking card games, alpha-beta search, computational complexity, machine learning, AI 1 Introduction A substantial part of artificial intelligence deals with investigating the extent to which machines are able to perform nontrivial complex human tasks. One of such tasks is playing games, a topic which over the years has received a lot of attention in artificial intelligence research [8,9,10], leading to a number of breakthroughs. A recent example is AlphaGo [21], where a combination of search algorithms and machine learning techniques is used to effectively beat humans at the highly 2 Jan N. van Rijn, Frank W. Takes, and Jonathan K. Vis complex game of Go. In general, a key problem in such games is that the search space of all possible game configurations is extremely large. This makes it difficult for algorithms to choose for example the next best move, whereas such a decision is often without much effort successfully taken by a human. In this paper we aim to explore this difference in competence of machines and humans, in particular in automatically assessing if a given instance of the game of Klaverjas can be won. Klaverjas is a trick-taking (card) game, played with the top eight cards from each suit of the French deck. Each card has a face f 2 F = f7; 8; 9; 10; J; Q; K; Ag, a suit s 2 S = f|; }; ~; ♠} and a rank r 2 R = (1;:::; 8). Cards with higher ranks are considered more powerful. One suit is designated the trump suit; cards from this suit are considered more powerful than all cards from other suits. Each card has a specific number of points associated to it, based on the rank and whether it is part of the trump suit or not. Table 1 displays for each card its value in points. Note that the rank of a card depends on the face value and whether it is part of the trump suit. Table 1. Rank and number of points per card. Rank Regular Trump 8 A 11 J 20 7 10 10 9 14 6 K 4 A 11 5 Q 3 10 10 4 J 2 K 4 3 9 0 Q 3 2 8 0 8 0 1 7 0 7 0 The game is played with four players that form two teams: one team consists of player N (north) and S (south) whereas the other consists of player E (east) and W (west). Each player starts with eight cards, to be played in each of the eight tricks. A trick is a part of the game in which each player plays one card. The first card of a trick can be freely chosen by the starting player. The other players must follow the leading card. If such a card is not available in a player's hand, a trump card must be played. Whenever a player must play a trump card, and a trump card has already been played in that trick, if possible, a higher rank trump card should be played. If the player can neither follow suit nor play a trump card, it is allowed to play any card that the player has left. It should be noted that there are also versions of the game in which always playing a (higher) trump card is not mandatory if a team mate has already played a trump card, referred to as \Amsterdams" rather than the version which we consider, which is \Rotterdams" Klaverjas. Once the fourth player has played his card, the trick has ended, and the winner of the trick is determined. If trump cards have been played, the trump Winning Hands in the Game of Klaverjas 3 card with the highest rank wins the tricks. If not, the card with the highest rank of the leading suit wins the trick. The player who played this card takes all the cards, his team receives the associated points and will start the next trick. The team that wins the last of the eight tricks is awarded 10 additional points. To win the game the team that started the game has to accumulate more points than the opposing team. If they fail to do so, i.e., they lose the game, which is referred to as nat, 162 points are awarded to the opposing team. Note that draws do not exist. When the starting team manages to win all eight tricks of the game they are awarded 100 bonus points, referred to as pit. Additionally, special meld points can be claimed by the team winning the trick when cards with adjacent face values are in the trick. These are for three ascending face values 20 meld points and for four ascending face values 50 meld points. For the King and Queen of trump, players can claim 20 meld points, in addition to other meld points already claimed in that trick. Finally, when four cards of the same face are played in the same trick, the team winning the trick can claim 100 meld points. For determining meld points, the order in which the players have played these cards is irrelevant. The addition of meld changes the dynamics of the game drastically, as players might sometimes be inclined to play a good card in a trick that is already lost, to prevent conceding meld points to the opposing team. Teams can choose not to claim meld points, for example when they already know they will lose the game. In this paper we consider the task of determining and predicting whether an instance of the game of Klaverjas is winning for a variant of the game, in which complete information on a team's cards is available. In addition, we assume that there is no bidding process of determining which player starts the game; the first player always starts and determines the trump suit. For this simplified version of the game, we consider the decision problem of, given a particular distribution of cards over players, determining whether this hand is winning for the starting player. We do so using both an exact algorithm based on αβ-search, as well as using a machine learning algorithm that based on feature construction mimics how a human decides whether a hand would be winning, for example based on counting high value cards. The results presented in this paper are useful for at least two types of new insights. First, in the real game, determining the starting player is done based on bidding, where players assess the quality of their dealt hand, based on whether they think they can win that hand. The approaches presented in this paper essentially perform this type of hand quality assessment. Second, as we employ both an exact approach and a machine learning approach, we can investigate the extent to which both are able to efficiently solve the game of Klaverjas. This will allow to investigate whether exact algorithms have the same difficulties with certain hands as an exact algorithm faces. The remainder of this paper is organized as follows. After discussing related work in Section 2, we introduce various definitions and necessary notation in Section 3. Then, an exact algorithm for solving a game of Klaverjas is presented in Section 4. Next, a machine learning approach is presented in Section 5. A 4 Jan N. van Rijn, Frank W. Takes, and Jonathan K. Vis comparison between the two is made in Section 6. Finally, Section 7 concludes the paper and provides suggestions for future work. 2 Related work Klaverjas is an example of the Jack-Nine card games, which are characterized as trick-taking games where the the Jack and nine of the trump suit are the highest-ranking trumps, and the tens and aces of other suits are the most valuable cards of these suits [16].

Load more